Learning system abstractions for human operators

This paper is concerned with the use of formal techniques for the analysis of human-machine interactions (HMI). The focus is on generating system abstractions for human operators. Such abstractions, once expressed in rigorous, formal notations, can be used for analysis or for user training. They should ideally be minimal in order to concisely capture the system behaviour. They should also contain enough information to allow full-control of the system. This work addresses the problem of automatically generating abstractions, based on formal descriptions of system behaviour. Previous work presented a bisimulation-based technique for constructing minimal full-control abstractions. This paper proposes an alternative approach based on the use of the L* learning algorithm. In particular, minimal abstractions are generated from learned three-valued deterministic finite-state automata. The learning-based approach is applied on a number of examples and compared to the bisimulation-based approach. The result of these comparisons is that there is no clear winner. However, the proposed approach has wider applicability since it can handle more types of systems than the bisimulation-based technique. Moreover, if no full-control abstraction can be generated due to a form of non-determinism in the system, the learning-based approach provides counterexamples that allow to detect and analyze that non-determinism. We also discuss how the well-known HMI issue of mode confusion can be analyzed through this approach.